zlxie 在 分享一个引导第三方DeepSeek-R1思考的system prompt 中发帖
众所周知DS的R1是现在体验最好的reasoning模型(能力强,CoT可见,擅长中文),但官方和siliconflow的服务目前都不太稳定,第三方服务容易跳过thinking过程。尽管DS官方文档指出
Avoid adding a system prompt; all instructions should be contained within the user prompt.
…
To ensure that the model engages in thorough reasoning, we recommend enforcing the model to initiate its response with “\n” at the beginning of every output.
但似乎绝大多数第三方API并未提供指定回复格式的功能,目前貌似只有together...